do I need to plan for GPU contention?

Maybe — it depends on your setup and workload.

You should plan for GPU contention if any of these are true:

• Multiple jobs/processes may share the same GPU
• You run in a multi-tenant environment (Kubernetes, shared servers, cloud instances with oversubscription)
• Your workload has spiky or unpredictable GPU demand
• You need consistent latency / throughput rather than “best effort”
• You’re using large models, training, or mixed inference + batch jobs

You may not need to worry much if:

• Each workload gets a dedicated GPU
• The machine is single-purpose
• You control scheduling tightly and know jobs won’t overlap

What to plan for:

• Admission control / scheduling: don’t launch more GPU work than you can serve
• Queueing and backpressure
• Isolation: one job per GPU, MIG/vGPU if applicable
• Monitoring: GPU utilization, memory use, latency, queue depth
• Fallback behavior: CPU fallback, reduced batch size, graceful degradation

If you tell me your context — e.g. training vs inference, single machine vs cluster, how many jobs/users, and whether you need latency guarantees — I can tell you how seriously to plan for it.